Repository navigation
OpenCode Go quota usage appears ~4x higher than displayed DeepSeek V4 Flash cost #42985
Description
Description
Possible OpenCode Go usage metering discrepancy with DeepSeek V4 Flash
I’m seeing a discrepancy between the cost shown in the usage graph and the Go quota consumption for deepseek-v4-flash.
On Aug 17, the usage graph shows:
deepseek-v4-flash (go): $3.31
At approximately the same time, my Go limits showed:
5-hour usage: 100%
Weekly usage: 44%
Monthly usage: 22%
According to the Go documentation, these limits correspond to $12 / $30 / $60 of usage respectively.
Therefore:
44% of $30 ≈ $13.20
22% of $60 ≈ $13.20
This is almost exactly 4x the $3.31 shown in the cost graph.
I was only using opencode-go/deepseek-v4-flash for this workload.
Could you confirm whether the Go quota meter uses a different internal cost multiplier than the cost shown in the usage graph for DeepSeek V4 Flash?
If not, there may be a roughly 4x discrepancy in quota accounting.
I can provide screenshots and detailed session/token usage logs if useful.
Plugins
No response
OpenCode version
No response
Steps to reproduce
No response
Screenshot and/or share link
No response
Operating System
No response
Terminal
No response
Activity
github-actions commentedon Aug 17, 2026
This issue might be a duplicate of existing issues. Please check:
If your issue is the same root cause, consider adding a comment there instead so the discussion stays in one place.
tnn226 commentedon Aug 17, 2026
One additional data point that may help identify the root cause:
my usage history shows $3.31, while weekly quota consumption was 44% of $30 = $13.20 and monthly was 22% of $60 = $13.20.
That is almost exactly a 4.0x discrepancy.
The workload used only deepseek-v4-flash (go).
xXStrayXx commentedon Aug 17, 2026
THIS IS A SCAM THEY MADE IT THAT CERTAIN MODELS HAVE 15$ QUOTA THIS IS A SCAM
jethrosalvador commentedon Aug 17, 2026
same problem here, I also use flash for 3.31 and it cost me my entire 5 hour rolling limit
n-osennij commentedon Aug 17, 2026
I agree. Except in my case, I lost the entire monthly limit, even though I had about 30% left in reserve.
pigletzzzzzz commentedon Aug 17, 2026
I'm having the same problem here. Last week when I used it, the cache hit rate was over 80%. Today the cache hit rate is below 10%. I'm using deepseek-v4-flash (go). I've used up my full 5-hour quota in just 5 minutes. Please find out what's wrong as soon as possible and provide a practical solution
sunxfancy commentedon Aug 17, 2026
Same problem, I've used up my full 5-hour quota in just 5 minutes
wulu007 commentedon Aug 17, 2026
same
zy25-debug commentedon Aug 17, 2026
I also encountered the same problem, running out of the 5-hour quota within just a few minutes.
Smankusors commentedon Aug 17, 2026
I believe it's not just 4x higher, but ~15x higher cost
Previously it was on double promotion, 63,300 requests per 5 hours
Now it's at 3,800 requests per 5 hours
https://opencode.ai/go
cwailam commentedon Aug 17, 2026
I'm having the same problem here.
1395139523-prog commentedon Aug 17, 2026
WegoW commentedon Aug 18, 2026
Since they have introduced models with 15/30 $ limits, the 60 $ monthly budget claim is completely misleading. If you max out a 15 $ model, your monthly usage is 100%.
Basically you have to multiply the costs of a model by the factor 60/Limit and sum it up. Then you get the real costs, that are charged on the 60$ budget.
Example:
Deepseek V4 Flash 15 $ Limit -> Factor 4
Mimo 2.5 60$ Limit -> Factor 1
Costs shown on usage dasboard:
Deepseek 10 $
Mimo 5 $
10 x 4 + 5 x 1 = 45 $
45/60 = 75%
ZH1K1NG commentedon Aug 18, 2026
Opencodego is integrated with Claude Code and Codex via ccSwitch, but during usage the cache hit rate is extremely low—already below 10%, and it keeps dropping with use, even trending towards nearly 0%!!!
Zaczero commentedon Aug 18, 2026
The 4× here is the quota factor for a $15 model, not a Flash-specific bug.
Flash’s Usage value in the table is what sets that factor (
60 / Usage). At $15 that is 4×:$3.31 × 4 = $13.24 — which is 44% of the $30 weekly cap and 22% of the $60 monthly cap. Those are the two percentages in the report.
Same rule as #41391, and the same arithmetic as #41146 and #43149.
One thing has moved since this was filed. The table at https://opencode.ai/docs/go/#usage-limits now lists Flash at Usage $30, which is a 2× factor. A clean 4× on 17 Aug means Flash was still $15 in billing that day, or the table has not caught up. The later 2.3× comments already match $30.
The “whole cap in five minutes” reports in this thread are a different problem. Those are cache reads falling over, which is #42935. The factor only multiplies whatever dollars that produces.
10 remaining items
CoderBuck commentedon Oct 20, 2026
New subscriber, first 5 hours, already locked out of every paid Go model. Posting here because my numbers reproduce the arithmetic in this thread exactly.
Account
TERM_PROGRAM=otty, zsh~/.config/opencode/plugins/otty-integration.jsConsole "Usage by Model" (4 models, 1 provider)
Total real spend: $3.47
Earlier capture, same table with only the two models that were used at the time:
Go card
Same card ~15 minutes earlier, before the limit was hit:
What happened
After 59
kimi-k3requests, every paid model —kimi-k3,deepseek-v4.1-flash,glm-5.3-flash— returns the same5-hour usage limit reachederror with the same reset countdown, including models whose own usage is still 0. Only the free models keep working.The arithmetic
Using the rule from this thread (
spend / that model's own window limit, summed across models):Matches the console exactly, and shows the mechanism in one line:
kimi-k3sits in the $15/mo tier, so its 5-hour cap is $3, i.e. 1/4 of the $12 the same five hours would allow for a $60-tier model. One model that is cheap in absolute terms (expensive in token prices: $3/M input, $15/M output) consumed 90.7% of the shared meter on its own.$3.47 of real spend — 5.8% of the documented $60 monthly allowance — and the account is fully blocked.
The documented rules say this should not happen
Read literally,
kimi-k3exhausting its own $3 five-hour cap should block onlykimi-k3, and a subscriber with $3.47 spent should still have $56.53 of monthly allowance left. The per-model request table on https://opencode.ai/go ("Kimi K3 — 110 requests / 5 hours", "31,580 requests / month" for GLM-5.3-Flash, …) is likewise unreachable, because whichever model reaches its own share first exhausts the pool for everything else.I subscribed based on those numbers. What is actually enforced is a single shared, normalized meter (
base × Σ(model usage / model limit)) with one aggregate gauge on the Go card, so the console gives no way to see which model drained the pool, and the session fails mid-work with no warning.I am not disputing that Go is metered, and I am happy to pay for what I use. I am asking for the documented behavior, or for the docs and the dashboard to state the real rule.
Requests
Screenshots are hosted in a public repo I control, since GitHub's attachment API is not available to CLI clients: https://github.com/CoderBuck/opencode-go-quota-evidence (no secrets in the images).
neriousy commentedon Oct 21, 2026
Thanks for the detailed numbers. This is how the Go allowance works rather than a metering bug. Go has one usage allowance shared across all models, and each model counts toward it at a different rate, set by its monthly limit in the docs. A model with a $15 monthly limit counts 4x as much as a $60 one for the same spend. When this was reported, DeepSeek V4 Flash had a $15 limit, which is exactly the 4x you calculated. The usage graph in the old console showed raw dollar cost, not the weighted amount, so the two looked inconsistent.
Limits have changed since: DeepSeek V4.1 Flash now has an $80 monthly limit, so it uses up the allowance much more slowly. The current per-model limits and estimated request counts are at https://opencode.ai/docs/go#usage-limits. If your usage still doesn't add up, email help@anoma.ly with your workspace ID.
msingelwa commentedon Oct 21, 2026
Any feedback on this?